[fragmede]

But the models can run tools, so wouldn't they just run the code, not get the expected output, and then exclude the bad code from their training data?

[bee_rider]

That seems like a feedback loop that’s unlikely to exist currently. I guess if intentionally plausible but bad data became a really serious problem, the loop could be created… maybe? Although it would be necessary to attribute a bit of code output back to the training data that lead to it.